Update dataset card to match 100% jihadv4 format & statistics

README.md

@@ -12,122 +12,27 @@ tags:
 - bangladesh-bank
 size_categories:
 - n<1K
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
-  - split: validation
-    path: data/validation-*
-- config_name: documents
-  data_files:
-  - split: train
-    path: documents/train-*
-dataset_info:
-- config_name: default
-  features:
-  - name: messages
-    list:
-    - name: content
-      list:
-      - name: text
-        dtype: string
-      - name: type
-        dtype: string
-    - name: role
-      dtype: string
-  - name: image
-    dtype: image
-  - name: task_type
-    dtype: string
-  - name: doc_type
-    dtype: string
-  - name: target_text
-    dtype: string
-  - name: source_file
-    dtype: string
-  - name: page_number
-    dtype: int64
-  - name: was_correct
-    dtype: bool
-  - name: batch_id
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 100024308
-    num_examples: 320
-  - name: validation
-    num_bytes: 25516410
-    num_examples: 80
-  download_size: 125400400
-  dataset_size: 125540718
-- config_name: documents
-  features:
-  - name: image
-    dtype: image
-  - name: doc_type
-    dtype: string
-  - name: was_correct
-    dtype: bool
-  - name: action
-    dtype: string
-  - name: confidence_score
-    dtype: float64
-  - name: latency_sec
-    dtype: float64
-  - name: target_json
-    dtype: string
-  - name: model_output
-    dtype: string
-  - name: corrected_output
-    dtype: string
-  - name: extraction_prompt
-    dtype: string
-  - name: classification_prompt
-    dtype: string
-  - name: timestamp
-    dtype: string
-  - name: source_file
-    dtype: string
-  - name: page_number
-    dtype: int64
-  - name: image_path
-    dtype: string
-  - name: has_dpo_pair
-    dtype: bool
-  - name: dpo_chosen
-    dtype: string
-  - name: dpo_rejected
-    dtype: string
-  - name: batch_id
-    dtype: string
-  - name: synced_at
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 63601796
-    num_examples: 200
-  download_size: 62528422
-  dataset_size: 63601796
 ---
 
-# 🏦 Bangladesh Bank Trade Finance IDP — Multi-Layout Fine-Tuning Dataset
+# 🏦 Bangladesh Bank Trade Finance IDP — Multi-User Collaborative Fine-Tuning Dataset
 
-Self-contained vision-language dataset for end-to-end extraction and classification of the **8 official Bangladesh Bank regulatory trade documents**.
-Generated with **100% diverse layouts, distinct company profiles, varying aesthetic themes, and zero repeated structures**.
-Strictly conforms to official Bangladesh Bank regulatory XML/JSON standards with zero hallucination guarantee (`MessageIdentifier == ""`).
+This dataset contains human-reviewed, verified, and corrected document extractions for the **8 official Bangladesh Bank regulatory trade-finance document types**.
 
-## 📊 Dataset Summary
+It is **completely self-contained** and structured for immediate Vision-Language Model (VLM) fine-tuning anytime from any environment (Colab, Kaggle, GPU cluster, or local), with built-in multi-annotator merge support and incremental delta uploads.
 
-| Split / Subset | Rows / Records |
-| :--- | :--- |
-| `documents` | 200 |
-| `train` | 320 |
-| `validation` | 80 |
+---
+
+## 📊 Dataset Structure & Splits
+
+| Split / Subset | Description | Examples |
+|---|---|---|
+| **`train`** | Ready-to-train multi-task VLM conversations (`messages` format) | 320 |
+| **`validation`** | Stratified held-out evaluation conversations for monitoring eval loss | 80 |
+| **`documents`** | Full document-level feedback records with model outputs, human corrections, confidence scores & DPO pairs (subset: `documents`) | 200 |
 
-### Document Types Distribution (200 Total Verified Documents)
-| Document Category | Count | Description |
-| :--- | :--- | :--- |
+### 📑 Document Types Represented
+| Document Type | Document Count |
+|---|---|
 | `air_waybill` | 25 |
 | `bill_of_entry` | 25 |
 | `commercial_lca` | 25 |
@@ -137,22 +42,29 @@ Strictly conforms to official Bangladesh Bank regulatory XML/JSON standards with
 | `industrial_lca` | 25 |
 | `ocean_bl` | 25 |
 
-- **Visual Diversity:** 8 Header styles (Full Banner, Split Header, Centered Crest, Minimal Barcode, Letterhead Box, Telex Mono, Sidebar Strip, Double Rule Hero), 5 Table formats (Zebra Stripes, Dense Grid, Minimal Underline, Floating Pills, Monospace Matrix), 4 Procedural Stamp configurations, and 12 distinct color palettes.
-- **Realistic Handwriting:** Natural handwriting pen fills and endorsements (Caveat, Kalam, Bradley Hand in varied inks: classic blue ballpoint, dark ink, royal blue, navy, blue-black) paired with 100% strictly validated schema JSON.
-- **Enterprise Profiles:** 50 distinct real Bangladeshi corporations & global exporters across pharma, steel, textiles, electronics, agro, FMCG, and heavy industry.
+**Total DPO Preference Pairs Available:** 10 (for Direct Preference Optimization)
+
+- **Visual Diversity:** 8 Header styles, 5 Table formats, 4 Procedural Stamp configurations, and 12 distinct color palettes.
+- **Realistic Handwriting:** Natural handwriting pen fills and endorsements in varied inks (blue ballpoint, dark ink, navy, royal blue) paired with 100% strictly validated schema JSON.
 - **Header Integrity:** `MessageIdentifier` is strictly excluded from document image layouts and is normalized as empty `""` in ground truth.
 
-## 🚀 Quickstart: Train in 10 Lines with Unsloth
+---
+
+## 🚀 Quickstart: Train in 10 Lines with Unsloth / TRL
+
+You can fine-tune Qwen3-VL / Qwen2.5-VL directly on this dataset without ANY manual data wrangling:
 
 ```python
-from unsloth import FastVisionModel
 from datasets import load_dataset
+from unsloth import FastVisionModel
 from trl import SFTTrainer, SFTConfig
 
 # 1. Load dataset directly from Hugging Face
-dataset = load_dataset("bisalsaha/bb-trade-idp-feedback")
+ds = load_dataset("bisalsaha/bb-trade-idp-feedback")
+train_data = ds["train"]
+eval_data = ds.get("validation")
 
-# 2. Load model with Vision LoRA
+# 2. Load model & attach vision LoRA adapters
 model, tokenizer = FastVisionModel.from_pretrained(
     "unsloth/Qwen3-VL-8B-Instruct-unsloth-bnb-4bit",
     load_in_4bit=True
@@ -164,11 +76,11 @@ model = FastVisionModel.get_peft_model(
     target_modules=["q_proj", "v_proj"]
 )
 
-# 3. Train
+# 3. SFT Trainer
 trainer = SFTTrainer(
     model=model,
-    train_dataset=dataset["train"],
-    eval_dataset=dataset["validation"],
+    train_dataset=train_data,
+    eval_dataset=eval_data,
     dataset_text_field="messages",
     max_seq_length=2048,
     args=SFTConfig(